One of the criteria that every camera review site uses in evaluating the quality of a new camera is image noise. As ISO increases noise normally increases as the sensor capture capabilities (or lack thereof) and boost processing electronics take their toll on the quality of the captured image.

There are many things that affect that noise, not the least of which is the size of the light-gathering pixels that capture the image. As discussed in detail in The Digital Sensor and The Digital Sensor Part 2, the size of the photosite is the reason the tiny sensors used in pocket point and shoot cameras are so limited in capabilities compared to the Digital SLR, which is where the major growth in the photographic market is now occurring.

The impact of photosite size is ably demonstrated in one of today's hottest DSLR cameras - the Nikon D3. This $5000 camera is a 12.2MP full-frame camera. Since it can also handle Nikon DX crop lenses and create a 5MP image at the common APS-C size seen in most of today's DSLR cameras, it is easy to understand that this 12MP full-frame has the same photosite size as a 5MP crop-sensor DSLR.

As discussed in detail in the Digital Sensor articles, the sensor is an analog light-gathering device. It is not a simple digital on-or-off digital device where size really matters very little. In a camera sensor the light gathering is analog and the larger the light gathering area, all else being equal, the better the light gathering ability.

Compared to today's 10 to 14.6MP sensors in top DSLR cameras with a 14x24 sensor, the Nikon D3 has huge photosites in its 24x36mm 12MP sensor. Combined with CMOS technology and evolved electronics the Nikon D3 has a normal ISO range to ISO 6400. As a comparison, the Canon XSI top sensitivity is 1600 and most prosumer models extend to ISO 3200.

As you can see in this dramatic ISO 6400 hockey shot produced by Dave Black for Nikon, D3 images at ISO 6400 are extremely low noise. Even more impressive is the ability to still capture usable images at ISO 12800 and all the way to ISO 25600 with much increased noise in a pinch.

Clearly noise is related to the size of pixels in analog sensors, but that is only part of the story. Often the manufacturer of the camera has made decisions, either informed or due to lack of expertise, to not process for noise at high ISO. Cameras that appear to exhibit much increased noise at high ISOs are sometimes capable of much lower noise. The poor noise performance is sometimes the result of decisions by the camera maker in how files will be converted to RAW or JPEG files in camera or in post-processing software.

The Sigma SD14 Wakeup Call

When Sigma SD14 on Vacation was published, one of the features discussed was photographs that tried to use the noise of the Sigma Foveon sensor at ISO 800 and even 400 as a creative tool. Like others, we assumed this noise issue at sensitivities that are a non-issue with other cameras is a problem of the Foveon sensor.

We received a very interesting email from a Gary Mercer, one of the professional photographers featured in the Popular Photography December 2007 Sigma SD14 promotion. He provided us with useful insight to the SD14 noise "problem" and a suggestion:

"I've been using Sigma DSLRs from the very humble beginning of the SD9 to the current SD14. There is a lot of controversy concerning the low light performance of the SD14 and also higher ISO 400 and up noise issue. I've found that other DSLR systems use in camera noise programs that produce soft, but less noisy images right out of the camera even when shooting raw. The supposedly raw images produced by most Bayer-sensor cameras are in fact heavily edited images prior to downloading them to the computer. The SD14 doesn't do any serious noise reduction in the image coming out of the camera. In fact, the SPP 3.2 and 2.5 versions of the Sigma Software barely address noise after the fact.

So how does someone who wants to shoot at higher ISOs with the SD14 deal with this? Imagenomic's Noiseware program is the best solution I've found to date. It is superb. If you sent me the raw images of the pictures you posted shot at ISO 800, I'm sure that Noiseware would take care of the noise in these images easily, salvaging them for you. I found this out after a helicopter shoot of the Hawaiian Islands with the SD14. We couldn't remove the doors off the helicopter, so I had to shoot through the canopy which lowered the light about two F-stops and also needed to use a polarizing filter which further lowered the light. It was a disaster in the making, forcing me to shoot at ISO 800. But after spending all this money on the helicopter, I was going to try to get my money's worth. When I got back to Florida and post processed the images, I thought the shoot was a complete bust, until a pro buddy of mine suggested Noiseware from Imagenomic. Even Noise Ninja couldn't clean up those images. Noiseware saved my shoot and I was able to save the images and use them in my latest gallery exhibition of my photography at www.pmgallery.info.

Moral of the story? The SD14 works just fine when you know the nature of the best and have the right tools to pull the results you want from this camera. It isn't super speedy, but I've got model test shots I've shot with it and lots of sample images on my SD14 test images website at pbase.com if someone wants to really see what is possible with the SD14 in real life."

Gary certainly has the credentials that persuaded us to try his suggestion and we were frankly astounded at the results we achieved by processing the Sigma JPEG files with Noiseware. The results were so impressive we tried the Noiseware processing with other camera images. Some cameras showed dramatic improvement in noise and others, like Nikon and Canon cameras, showed very little improvement with Noiseware processing. A large selection of images has been Noiseware-processed in the last few weeks. With this experience with a wide range of digital images from a cross-section of cameras, it is time to discuss the impact of Noiseware in more detail.

How does Noise Reduction Work?

If you have ever looked closely at the lower noise images from a Nikon or Canon camera, you will notice that lower noise at high ISO is something of a balancing act. Low noise is normally achieved at the expense of loss of detail in high ISO images. The low-noise appearance is the result of smearing and softening details and sharpness, and the effectiveness of the noise reduction is really very subjective. It is a balancing act between sharpness on the one hand and noise on the other. It is even more complicated than this as noise reduction can be applied in both Chroma and Luminance dimensions and the impact of each is different. However, in the end noise reduction is always a balance between loss of detail and visible noise.

There have been many software noise reduction software programs in recent years, but most have left us with the impression that it was just as easy to live with the noise as it was with the soft smeared images of noise reduction software. Perhaps this is the result of viewing the effects of noise reduction software on Canon and Nikon digital images. As some review sites point out endlessly, Canon and Nikon understand high ISO noise and they do a good job with in-camera JPEG processing of reaching a compromise between noise and sharpness. Conversely, these same review sites endlessly point out that Sigma, Pentax, and Olympus (until recently) have noise problems and don't correctly process images. Olympus has apparently seen the noise processing light since recent Olympus model JPEG images are actually now sharper and lower noise at high ISO speeds than the RAW images captured under the same conditions. We have seen this ourselves, and so have several other respected camera review sites.

As explored in the Digital Sensor, one great appeal of the CMOS sensor over the CCD is the ability to combine noise reduction and other processing on the sensor itself. Canon has enjoyed this advantage exclusively until recently, and Canon has also been widely praised for the exceptional low-noise of their high ISO images compared to every other manufacturer in that period. The fact is Canon has used noise reduction in processing RAW and JPEG files on their image sensors. Now Sony and Nikon are doing the same in the D300 and A700 model CMOS sensor.

Sony engineers had to be scratching their head when they saw recent reviews of the A700 that screamed about noise-processing in their RAW images. The Nikon D300 is the same sensor that has noise reduction built onto the sensor just as Canon CMOS sensors do. We suppose Sony got the temper tantrum because they weren't Nikon, and because they were not as subtle as Nikon in their discussion of sensor technology. Also, in fairness to the noise reduction complaints in some reviews, the supporting Sony BIONZ processing was not very subtle compared to the EXSPEED circuitry used by Nikon. Sony has subsequently refined their post processing with firmware upgrades to the A700.

Regardless, the fact remains that Canon, Nikon, Sony, and Olympus all likely employ noise reduction in the creation of their RAW files with their CMOS sensors, and certainly they do so in the conversion to JPEG files. Sigma uses almost no noise reduction in their file conversion, and Pentax uses very little noise reduction in their in-camera JPEG processing, which is why Pentax reviews always suggest shooting RAW. The Pentax image-processing software (Silkypix) and Photoshop RAW processing certainly do employ more noise reduction than Pentax does in-camera, and thus yield better appearing images at high ISO.

Noiseware Software Noise Reduction

Noiseware is a software program from Imagenomic. The program claims to take a different approach to noise reduction by using Artificial Intelligence techniques for noise reduction instead of simple median filters. The core of the system is an automatic profile system that attempts to analyze and recognize noise patterns for more effective noise detection. The program is adaptive and uses EXIF data and the results of its analysis to build a noise processing algorithm that becomes more effective as more images are processed. In automatic mode the program considers the image as a whole. Noiseware also features the option of a manual mode for manual noise profiling by specifying certain regions for noise reduction.

Noiseware is available as a standalone program, or as a plug-in for Adobe Photoshop or Elements. Current versions are available for the MAC or PC, with current PC versions working under Vista or XP. The cost is free for the Community Edition standalone with automatic profiling and manual adjustments but no adaptive AI learning. The Standard Edition standalone is $29.95 and Professional Edition standalone is $49.95. Free trial downloads are available for all versions in the Try Before Buy format.

Noiseware Standalone Versions
	Community Edition	Standard Edition	Professional Edition
Processing
Image processing format	JPEG, PNG, BMP, TIFF	JPEG, PNG, BMP, TIFF	JPEG, PNG, BMP, TIFF
Embedded ICC support	-	-	X
Batch processing	-	X	X
Custom filter preset support	-	-	X
Noise Profile
Automatic noise profiling	X	X	X
Self-learning noise profile builder	-	X	X
Opening/Saving
Open JPEG, PNG, BMP, TIFF (24 and 48 bpp)	X	X	X
Save result as JPEG	X	X	X
Save result as TIFF and PNG (24 bpp)	-	X	X
Save result as 48(16) bpp TIFF	-	-	X
Drag-n-Drop from Windows Explorer	X	X	X
Copy To & Paste From Clipboard	-	X	X
Preserving image EXIF data	-	X	X
Availability	Free Download	15-day Trial or Buy USD 29.95	15-day Trial or Buy USD 49.95

As an alternative, Noiseware is available as a plug-in for Adobe Photoshop or Elements at $49.95 for the Standard Plug-In and $69.95 for the Professional version.

Noiseware Plug-In Versions
	Standard Plug-in	Professional Plug-in
Image processing format	Any format supported by Photoshop
Embedded ICC support	via Photoshop
Photoshop Action support	-	X
Batch processing	-	via Photoshop Actions
RGB	X	X
LAB	Single Channel only	X
CMYK, Multichannel	Single Channel only	Single Channel only
8 bits/channel	X	X
16 bits/channel	-	X
Smart Filter	-	X
Custom filter presets	X	X
Automatic Noise Profiling	X	X
Self-learning Noise Profiles	X	X
Manual Noise Profiling	X	X
Multiple Previews	X	X
Bracketing	X	X
Operating system	Windows 2000/XP/Vista Mac OS X 10.3.x/10.4.x/10.5.x (PowerPC/Intel)
Compatible host program	Adobe Photoshop 7.0, CS, CS2 and CS3 Adobe Photoshop Elements 2/3/4/5/6 Corel Paint Shop Pro 9 and X Corel Draw 10, 11 and 12 Microsoft Digital Image Suite 2006 Ulead PhotoImpact XL, 10 and 11
Availability	Download or Buy USD 49.95	Download or Buy USD 69.95

Those who want both the Plug-In and the Standalone versions can get the Professional Bundle for $79.90. With the Community Edition offering free licensing and all other versions offering Try-It-First downloads, you can easily find a Noiseware version to match your noise processing needs.

All images on the next few pages were processed with the free Community Edition of Noiseware using automatic processing. Results should be considered baseline as even more effective processing is available with adaptive processing and manual tweaking. The goal here is not to show everything that can be done with a noise reduction program like Noiseware. Rather, we are trying to show the impact of even the lowest level Noiseware processing on images that are considered high ISO noise problems in many reviews. This is to show you what anybody can do to effectively control noise.

Sigma SD14 at ISO 800

This all began with suggestions on how high noise seen in Sigma SD14 ISO 800 images could be tamed and controlled. Our first look at the impact of Noiseware is Sigma SD14 ISO images first published in Sigma SD14 on Vacation. As stated, all images were processed with just the free Community Edition of Noiseware using just auto processing. You can do an even more effective job in some cases by tweaking the noise reduction parameters manually in the Community Edition. The real power of Noiseware comes in the AI-enabled Standard and Pro versions, which get more effective the more they are used. These are baseline conversions just to show you that the impact of Noiseware on the SD14 is truly significant.

In each pair of images, the top images are JPEGs that were processed from RAW in Sigma Photo Pro 2.5 and saved as highest quality JPEG images. The matching image just below is the same image that has been auto processed in Noiseware Community Edition.

Sigma SD14 at lower ISO

You would expect the greatest improvement in Noiseware-processing of SD14 images to be at the higher ISO values. However, those who shoot only at ISO 100 with the Foveon may be surprised to see there is also an improvement in quality at lower ISO. This makes ISO 200 and 400 useful for even the most demanding Foveon shooters. Most of the images below were captured at ISO 400. The two Macro shots were both captured at ISO 200.

Sigma SD14 Test Sequence

Having seen the dramatic noise reduction possible with SD14 images using Noiseware, the ISO crop test for the SD14 was run again, with both crops and full images processed with Noiseware. Individual images processed in times from less than a second for smaller images to 4 seconds for the largest in this sample. Batch processing is also supported in the Standard and Pro versions, which are definitely recommended if you plan to make a habit of processing your pictures.

Sigma SD14 Photo Pro 2.5 JPEG vs. Noiseware Processed
ISO	SD14 JPEG Captured RAW, Photo Pro JPEG	SD14 JPEG Additional Noiseware Processing
50
100
200
400
800
1600
Click on any of the above image crops for the full image. Note: Full size images are between 0.6MB and 2.9MB!

Pentax K20D

The 14.6MP Pentax K20D is the highest resolution sensor available in a prosumer camera today. There are pro models at $8000 or more that feature even higher resolutions, but the $1299 MSRP K20D is a long way from that price class. When it was introduced in late 2007, the K20D was the first 14MP class camera, and its introduction surprised the photo establishment with a Pentax-designed and Samsung-manufactured CMOS sensor.

Tests of K20D resolution show the highest resolution available in its class at lower ISOs. However, resolution and image quality at higher ISO settings are compromised by the higher noise of the K20D at these settings. It appears Pentax made the conscious decision to use minimum in-camera noise processing at higher ISOs.

Pentax K20D In-Camera JPEG vs. Noiseware Processed
ISO	Pentax K20D In-Camera JPEG	Pentax K20D Noiseware Processed
100
200
400
800
1600
3200
6400
Click on any of the above image crops for the full image. Note: Full size images are between 1.9MB and 11.7MB!

Noiseware completely changes the dynamics of high ISO image capture on the K20D. The comparison of JPEG images that have been Noiseware-processed to the original in-camera JPEG images shows a dramatic improvement in image quality, particularly at ISO 800 and beyond.

Sony A350

Several months ago, Sony introduced the mid-level A350 with a 14.2MP CCD sensor. The Sony CCD is unusual in that it is the only recent new sensor introduction that is not based on CMOS technology. The A350 is aimed at the top of the entry-level market and Sony targets the A700 at the prosumer/serious amateur market.

In testing the A350 it is clear that Sony is already performing quite a bit of in-camera noise processing with JPEGs produced by the A350. This is very similar to the in-camera processing performed by Canon and Nikon, although a book could be written on the subtle differences in "noise reduction philosophy" among these three companies.

Even with significant in-camera processing Noiseware can improve image quality at higher ISO settings, making those higher speeds more useful for the photographer using an A350. Improvements are much more subtle than the Sigma SD14 and Pentax K20D improvements but they can still be clearly seen.

Sony A350 In-Camera JPEG vs. Noiseware Processed
ISO	Sony A350 In-Camera JPEG	Sony A350 Noiseware Processed
100
200
400
800
1600
3200
Click on any of the above image crops for the full image. Note: Full size images are between 1.6MB and 6.2MB!

Our Take

While we began with exploring the impact of Noiseware on Sigma Foveon images, we found this program was also very effective with Pentax K20D and Sony A350 high ISO images. The greatest improvements seem to be in those images that have the lowest amount of in-camera noise processing.

The image improvement is pronounced in Sigma SD14 images that are reported to have little or no noise reduction applied in the Sigma Photo Pro software. Noiseware was also quite effective in processing Pentax K20D images from their Samsung 14.6MP sensor. Pentax applies smaller amounts of noise reduction in their image processing so the addition of Noiseware can have a substantial impact on the quality of high ISO images.

The Sony A350 is already employing a substantial amount of noise reduction in its in-camera processing of high ISO images. Noiseware can also improve these images, but the results are more subtle than those seen with less processed JPEGs. Improvements are certainly visible but they are not the night day results sometimes seen with Sigma SD14 and Pentax K20D images.

Canon and Nikon JPEG images processed with Noiseware behave much like the Sony A350 images. You will see improvements, but they are not as dramatic as some other images. This would certainly imply that both Canon and Nikon are employing large amounts of in-camera noise reduction, just as Sony does, which is contrary to what many assume. While Nikon and Canon seem to employ a similar amount of noise-reduction, this is not to say they all process noise in the same manner. A book could be written on the "noise reduction philosophies" of the three companies because they appear to be making different choices with their noise-reduction processing. However, many incorrectly criticize Sony, for instance, on their "heavy" image processing in the A350. Canon and Nikon also heavily process images in-camera; they just make slightly different choices in their processing algorithms.

The "selective" processing we saw with Noiseware certainly supports their claim that Noiseware is not "median-based" like most other noise reduction software. The program seems most effective where noise is the highest and it has the admirable trait of not over-processing images that need subtle noise fixes. In processing images for this review, Noiseware did live up to their claim of reducing noise without overly reducing sharpness in the image. All of this was with the free Community Edition. You can expect the licensed AI editions of the standalone and plug-in editions of Noiseware to be even better at selective processing, and to actually improve in effectiveness with extended use.

Noiseware only works with files that have been processed in-camera or with another software program like Photoshop. The standalone requires JPEG, PNG, BMP, or TIFF files for processing. The plug-in version can process any file format the native program supports, but Camera RAW is also a Photoshop plug-in. That means you can't use Noiseware with a RAW image as it is designed for post-processing.

With Noiseware used for post-processing images, the Sigma SD14 becomes a completely different camera. You will no longer be afraid to shoot at ISO 800 and even ISO 1600 is usable in a pinch, though it is not the same low-noise as ISO 1600 on competing cameras. Noiseware is also extremely effective in post-processing of Pentax K20D high ISO images.

The impact of Noiseware on cameras that already employ heavy noise reduction in-camera is not as dramatic. Sony A350 images are improved, but the effect is more subtle. The same is true of Canon and Nikon images which have been processed in-camera as JPEG images. These images are improved but changes are more subtle.

This selective behavior of Noiseware makes it an extremely useful noise reduction program. It works best where it is needed most in high ISO noise images such as those found with the Sigma SD14 and Pentax K20D. Noiseware does much less to images that have already experienced substantial noise reduction such as Sony A350, Nikon, and Canon images. It also seems to have little impact on the sharpness of these previously processed images, which is certainly a good thing.